Ungreedy Methods for Chinese Deterministic Dependency Parsing

Abstract

Deterministic dependency parsing has often been regarded as an efficient algorithm while its parsing accuracy is a little lower than the best results reported by more complex methods. In this paper, we compare deterministic dependency parsers with complex parsing methods such as generative and discriminative parsers on the standard data set of Penn Chinese Treebank. The results show that, for Chinese dependency parsing, deterministic parsers outperform generative and discriminative parsers. Furthermore, basing on the observation that deterministic parsing is a greedy algorithm which chooses the most probable parsing action at every step, we propose three kinds of ungreedy deterministic dependency parsing algorithms to globally model parsing actions. We take the original deterministic parsers as baseline systems. Results show that ungreedy deterministic dependency parsers perform better than the baseline systems while maintaining the same time complexity, and our best result improve much over baseline.